Skip to content
/

Claude Conductor (rbarcante)

claude-conductor-rbarcante · rbarcante/claude-conductor · ★ 53 · last commit 2026-03-31

Context-Driven Development plugin that enforces a spec-plan-implement lifecycle with Python-CLI context management, pattern library, and ADR logging for Claude Code.

Best whenEvery implementation decision should be documented at the time it is made — decisions.md as a first-class artifact prevents future context loss.
Skip ifCoding before speccing (spec.md must precede implement), Skipping plan approval gate
vs seeds
claude-conductor(Archetype 4 — markdown scaffold, 0/0/0/0 primitives), but rbarcante adds 9 commands, 6 sub-agents, 10 protocols, a Pyth…
Primitive shape 22 total
Commands 9 Skills 7 Subagents 6
00

Summary

Claude Conductor (rbarcante) — Summary

Claude Conductor by Ricardo Barcante is a full-featured Claude Code plugin that implements a Context-Driven Development (CDD) methodology with a spec-plan-implement lifecycle. Unlike the seed superbasicstudio/claude-conductor which provides only static markdown templates with no orchestration, this version ships 9 slash-commands, 6 sub-agents, 10 protocol files, a Python CLI helper (conductor_cli.py), and a Pattern Reference Layer of reusable best-practice patterns. The core workflow is: setup project context → create a "track" (feature/bugfix) with generated spec.md and plan.md → execute against the plan with TDD enforcement → auto-trigger code review on completion. A Universal File Resolution Protocol (UFRP) and AI-Optimized Templates with dual-format pattern docs make it one of the most elaborately documented CDD plugins in this corpus. It stores all state in the conductor/ directory as structured markdown files with a metadata.json per track. Compared to the seed claude-conductor (Archetype 4 — markdown scaffold), this belongs closer to Archetype 1 (skills-behavioral) but with explicit slash commands, making it a hybrid CDD plugin framework with real orchestration semantics, closest in spirit to lackeyjb/claude-conductor (which it predates as the upstream source) but significantly more featureful with its Python CLI, Pattern Reference Layer, Skill Ecosystem, and ADR logging.

01

Overview

Claude Conductor (rbarcante) — Overview

Origin

Philosophy

The tagline is "Measure twice, code once." Conductor treats context as a managed artifact co-equal with code. The core doctrine is:

"By treating context as a managed artifact alongside your code, you transform your repository into a single source of truth that drives every agent interaction with deep, persistent project awareness."

The plugin explicitly warns users about token consumption upfront — it reads and analyzes project context, specifications, and plans — flagging this as a known cost of the methodology.

Key design beliefs

  1. Context > Code: Every agent interaction is driven by stored, structured context files (product.md, tech-stack.md, workflow.md) rather than re-elicited on each prompt.
  2. Tracks as unit of work: All development is organized into tracks (features/bugs/refactors/docs) with a spec.md defining what and a plan.md defining how.
  3. Brownfield-first: Unlike the lackeyjb variant, rbarcante explicitly supports brownfield project analysis with codebase pattern detection and generates conductor/docs/ pattern documentation for existing codebases.
  4. Extensible skill ecosystem: A skill-registry.json and manifest schema allow third-party skills (e.g., acli-jira, glab-cli, java-best-practices, typescript-best-practices) to be plugged into the workflow.
  5. Decision logging via ADR: Architecture Decision Records are captured in conductor/tracks/<track_id>/decisions.md to explain the "why" behind implementation choices.
  6. Quality Intelligence: Automated anti-pattern detection and test coverage analysis during implementation with actionable suggestions.

Cross-references

The seed superbasicstudio/claude-conductor lists rbarcante as a derivative fork. The rbarcante version adds substantially to the methodology with the Python CLI, skill ecosystem, protocols layer, and UFRP — making it an independent enhancement rather than a cosmetic fork.

02

Architecture

Claude Conductor (rbarcante) — Architecture

Distribution

  • Type: Claude Code plugin (marketplace-installable)
  • Install method:
    /plugin marketplace add rbarcante/claude-conductor
    /plugin install conductor@claude-conductor
    
    or manual: cp -r claude-conductor ~/.claude/plugins/conductor
  • Required runtime: Python (for conductor_cli.py), Claude Code

Directory structure

rbarcante/claude-conductor/
├── .claude-plugin/
│   ├── plugin.json           # Plugin manifest (name: conductor, v1.6.0)
│   └── marketplace.json      # Marketplace metadata
├── commands/                 # 9 slash commands
│   ├── setup.md
│   ├── newTrack.md
│   ├── implement.md
│   ├── status.md
│   ├── revert.md
│   ├── codeReview.md
│   ├── patterns.md
│   ├── skills.md
│   └── snippet.md
├── agents/                   # 6 sub-agents
│   ├── code-quality-analyzer.md
│   ├── codebase-pattern-detector.md
│   ├── git-history-analyst.md
│   ├── security-scanner.md
│   ├── test-coverage-analyzer.md
│   └── track-context-researcher.md
├── protocols/                # 10 protocol files
│   ├── ai-template-generation.md
│   ├── codebase-analysis.md
│   ├── coverage-intelligence.md
│   ├── decision-capture.md
│   ├── git-isolation.md
│   ├── pattern-resolution.md
│   ├── quality-analysis.md
│   ├── skill-loading.md
│   ├── stack-detection.md
│   └── verify-setup.md
├── skills/                   # Skill ecosystem
│   ├── skill-registry.json
│   ├── skill-registry-schema.json
│   ├── manifest-schema.json
│   ├── typescript-best-practices/
│   ├── testing-strategies/
│   ├── api-design/
│   ├── java-best-practices/
│   ├── acli-jira/
│   ├── glab-cli/
│   └── conductor-methodology/
├── scripts/
│   ├── conductor_cli.py      # Python CLI helper
│   ├── commands/             # CLI subcommands
│   └── lib/                  # Library code
├── patterns/                 # Pattern Reference Layer
│   └── index.md
├── protocols/                # Additional protocols
├── snippets/                 # Code snippet library
│   └── index.md
├── templates/                # AI-Optimized Templates
├── docs/                     # Documentation
└── conductor/                # Project runtime dir (written at setup)
    ├── product.md
    ├── tech-stack.md
    ├── workflow.md
    ├── tracks.md
    └── tracks/<track_id>/
        ├── spec.md
        ├── plan.md
        ├── decisions.md
        ├── metadata.json
        ├── review.md
        └── index.md

Target AI tools

  • Primary: Claude Code
  • Claude Code version: >=0.1.0 (per plugin.json)

Languages / runtimes

  • Plugin prompt files: Markdown
  • CLI helper: Python
  • No Node.js or npm required for the plugin itself
03

Components

Claude Conductor (rbarcante) — Components

Commands (9)

Command Purpose
/conductor:setup Scaffold project context, detect brownfield vs greenfield, generate all context files
/conductor:newTrack Create a new feature/bugfix track: generate spec.md + plan.md + decisions.md
/conductor:implement Execute tasks from the current track's plan with TDD enforcement
/conductor:status Display overall and per-track progress
/conductor:revert Git-aware revert at track/phase/task granularity
/conductor:codeReview Comprehensive code review (also auto-triggered at track completion)
/conductor:patterns Browse/search the Pattern Reference Layer
/conductor:skills Manage Conductor skills (list, info, enable, disable)
/conductor:snippet Browse/search/display code snippets from Snippet Library

Sub-agents (6)

Agent Purpose
code-quality-analyzer Automated quality analysis and anti-pattern detection
codebase-pattern-detector Detect and document existing code patterns (brownfield)
git-history-analyst Analyze git history for revert operations
security-scanner Scan code for security issues
test-coverage-analyzer Coverage intelligence and test gap detection
track-context-researcher Research and consolidate track context

Protocols (10 internal reference files)

Protocol Purpose
ai-template-generation How to generate AI-Optimized Templates
codebase-analysis Brownfield codebase analysis process
coverage-intelligence Test coverage analysis protocol
decision-capture ADR logging methodology
git-isolation Git branch/worktree management
pattern-resolution Universal File Resolution Protocol (UFRP)
quality-analysis Code quality gate protocol
skill-loading How to load and activate skills
stack-detection Tech stack auto-detection rules
verify-setup Pre-flight checks for commands

Skills (bundled in skill ecosystem)

Skill Purpose
typescript-best-practices TypeScript coding standards
testing-strategies Testing approach guidance
api-design API design patterns
java-best-practices Java coding standards
acli-jira Jira CLI integration
glab-cli GitLab CLI integration
conductor-methodology Core CDD methodology skill

Scripts

Script Purpose
scripts/conductor_cli.py Main Python CLI helper
scripts/commands/ CLI subcommand implementations
scripts/lib/ Shared library code

Templates

  • Dual-format AI-Optimized Templates with AI Quick Reference sections
  • templates/workflow.md — customizable workflow template
  • templates/askuserquestion-patterns.md — patterns for interactive prompts

Snippets

  • snippets/index.md — searchable code snippet library index

Patterns

  • patterns/index.md — Pattern Reference Layer; surfaced during implementation
05

Prompts

Claude Conductor (rbarcante) — Prompts

Excerpt 1: commands/newTrack.md — System Directive with Plan Mode protocol

<system_directive>

## 1.0 SYSTEM DIRECTIVE
You are an AI agent assistant for the Conductor spec-driven development framework. Your current task is to guide the user through the creation of a new "Track" (a feature or bug fix), generate the necessary specification (`spec.md`) and plan (`plan.md`) files, and organize them within a dedicated track directory.

This command uses a **two-phase workflow** powered by Claude Code's Plan Mode:
- **Phase A (Plan Mode):** Research, compose spec + plan content in read-only mode, write to CC plan file for review
- **Phase B (Normal Mode):** After user approval, create branch, scaffold directory, write files, register, commit

<note type="critical">
You must validate the success of every tool call. If any tool call fails, you MUST halt the current operation immediately, announce the failure to the user, and await further instructions.
</note>

</system_directive>

Technique: XML-tagged directive sections with explicit tool-call validation. Uses Claude Code's Plan Mode as a two-phase approval gate. Critical notes in typed XML tags.


Excerpt 2: commands/implement.md — CLI injection + warm-start detection

# Context

!`python ${CLAUDE_PLUGIN_ROOT}/scripts/conductor_cli.py --json implement parse-tracks`

## 1.0 SYSTEM DIRECTIVE

You are an AI agent for the Conductor spec-driven development framework. Your task is to implement a track. Follow this protocol precisely.

CRITICAL: Validate every tool call. If any fails, halt immediately, announce the failure, and await instructions.

---

## CLI Reference

```bash
# Status & archiving
python ${CLAUDE_PLUGIN_ROOT}/scripts/conductor_cli.py implement update-status TRACK_ID STATUS
python ${CLAUDE_PLUGIN_ROOT}/scripts/conductor_cli.py implement archive TRACK_ID
# Consolidated context (replaces multiple Reads)
python ${CLAUDE_PLUGIN_ROOT}/scripts/conductor_cli.py --json tracks read-context TRACK_ID
python ${CLAUDE_PLUGIN_ROOT}/scripts/conductor_cli.py --json implement git-snapshot [--diff-stat-only]

1.0.1 WARM START DETECTION

PROTOCOL: Detect if implement was invoked directly from newTrack (context already loaded).

  1. Check arguments for --warm-start flag
  2. If --warm-start is present:
    • Announce: "⚡ Warm start mode — skipping redundant setup (context carried over from newTrack)."

**Technique**: Context injection at prompt head via `!`-prefixed bash command (live CLI data). JSON output passed as structured context. Fallback instructions if CLI fails. Warm-start optimization for command chaining.

---

## Prompting techniques observed

1. **Live context injection** (`!command` syntax) — tracks are injected as JSON at prompt head, not hardcoded
2. **CLI-as-tool pattern** — Python CLI replaces many Read calls; structured `--json` output feeds into agent decisions
3. **Typed XML sections** — `<system_directive>`, `<note type="critical">`, `<protocol name="...">`, `<cli_reference>` for clear structural parsing
4. **Fallback instructions** — every CLI call has a fallback if the tool fails
5. **Tool allowlist** per command — explicit `allowed-tools:` frontmatter restricts what each command can invoke
6. **AskUserQuestion protocol** — standardized interactive prompt pattern across all commands
09

Uniqueness

Claude Conductor (rbarcante) — Uniqueness

differs_from_seeds

The closest seed is the superbasicstudio/claude-conductor (Archetype 4 — markdown scaffold), but rbarcante diverges fundamentally: it adds 9 slash-commands, 6 sub-agents, 10 protocol files, a Python CLI helper with 15+ subcommands, a skill ecosystem with registry, ADR decision logging, a Pattern Reference Layer, and a Snippet Library. Where superbasicstudio provides only static templates that Claude passively reads, rbarcante provides an active orchestration system with Plan Mode gating, warm-start optimization, and brownfield codebase analysis. The methodology name "Context-Driven Development" is shared, but rbarcante is to superbasicstudio what spec-kit is to agent-os — the operational implementation vs. the scaffolded docs. Among non-seed variants, rbarcante is the upstream inspiration for lackeyjb (which simplifies it) and for MadAppGang's conductor plugin (which embeds it in a larger multi-plugin marketplace).

Positioning

  • Versus lackeyjb: rbarcante is richer (Python CLI, Pattern Layer, UFRP, Snippet Library, ADRs, auto code review, brownfield detection, 10 protocol files vs lackeyjb's 0); lackeyjb is simpler and cleaner
  • Versus MadAppGang conductor: rbarcante is standalone; MadAppGang embeds a similar but slimmed-down conductor plugin in a 14-plugin marketplace
  • Versus spec-driver (seed): both have skills + sub-agents; spec-driver has 24 skills/0 commands; rbarcante has 9 commands/6 agents/0 auto-activating skills
  • Versus taskmaster-ai (seed): taskmaster has MCP + tasks.json; rbarcante has file-based markdown state + Python CLI

Distinctive properties

  1. UFRP (Universal File Resolution Protocol) — flexible file organization with dynamic path resolution via index files
  2. Pattern Reference Layer — automatically surfaced during implementation based on task context
  3. ADR logging — captures the "why" behind implementation choices in decisions.md
  4. AI-Optimized Templates — dual-format patterns with AI Quick Reference sections
  5. Skill Ecosystemskill-registry.json + manifest schema for extensible domain-specific guidance
  6. Warm-start optimization--warm-start flag chains newTrack directly to implement without redundant setup

Observable failure modes

  1. Token consumption: Explicitly warned in README — reading all context files at every command adds latency and cost
  2. Python dependency: conductor_cli.py requires Python in PATH; if missing, CLI calls fail and fallback instructions must be followed manually
  3. Over-engineering risk: 9 commands + 10 protocols + 6 agents + pattern layer may overwhelm small projects or solo developers
  4. Track sprawl: No automatic archival trigger; stale tracks accumulate in conductor/tracks/
04

Workflow

Claude Conductor (rbarcante) — Workflow

Phases

Phase Command Artifacts produced
1. Project Setup /conductor:setup conductor/product.md, conductor/product-guidelines.md, conductor/tech-stack.md, conductor/workflow.md, conductor/tracks.md, conductor/index.md, conductor/code_styleguides/, conductor/docs/ (brownfield)
2. Track Creation /conductor:newTrack conductor/tracks/<id>/spec.md, conductor/tracks/<id>/plan.md, conductor/tracks/<id>/decisions.md, conductor/tracks/<id>/metadata.json, conductor/tracks/<id>/index.md
3. Implementation /conductor:implement Updated conductor/tracks/<id>/plan.md (task checkboxes), conductor/tracks.md (status), conductor/tracks/<id>/review.md (on completion)
4. Code Review /conductor:codeReview (auto or manual) conductor/tracks/<id>/review.md
5. Status check /conductor:status (read-only, no artifacts)
6. Revert (optional) /conductor:revert Git history modification

Approval gates

  1. Plan review before implement: User must review and approve plan.md before /conductor:implement proceeds.
  2. Phase verification: At the end of each implementation phase, user is guided through manual verification.
  3. Code review gate: Auto-triggered code review on track completion; non-blocking (user can skip or proceed regardless of findings).
  4. AskUserQuestion tool: All interactive prompts use the AskUserQuestion tool protocol — structured choice lists, not free-form.

TDD enforcement

The workflow enforces TDD when workflow.md specifies it:

  1. Write failing test (Red)
  2. Write minimum code to pass (Green)
  3. Refactor while tests pass (Refactor)

Track types

feature | bugfix | refactor | docs | chore

Plan Mode

/conductor:newTrack uses Claude Code's Plan Mode:

  • Phase A (Plan Mode, read-only): Research, compose spec + plan content, write to CC plan file for review
  • Phase B (Normal Mode): After user approval, create branch, scaffold directory, write files, register, commit

Git integration

  • Branch created per track (via suggest-branch CLI command)
  • Commits linked to track/phase/task via conductor_cli.py
  • revert analyzes git history to undo logical work units

Warm-start protocol

/conductor:implement detects if invoked directly from newTrack (--warm-start flag) and skips redundant setup steps.

06

Memory Context

Claude Conductor (rbarcante) — Memory & Context

State storage

All state is file-based in the conductor/ directory created at setup:

File Contents
conductor/product.md Product vision, goals, user personas
conductor/product-guidelines.md Brand voice, standards, codebase patterns (brownfield)
conductor/tech-stack.md Language, framework, database, tooling preferences
conductor/workflow.md Development methodology (TDD, commit strategy)
conductor/tracks.md Master index of all tracks with status
conductor/index.md Navigation index
conductor/settings.json Plugin settings, enabled skills
conductor/code_styleguides/ Language-specific style guides
conductor/docs/ Pattern documentation (brownfield projects)
conductor/tracks/<id>/spec.md Feature specification
conductor/tracks/<id>/plan.md Implementation plan with task checkboxes
conductor/tracks/<id>/decisions.md Architecture Decision Records (ADRs)
conductor/tracks/<id>/metadata.json Track state, type, timestamps
conductor/tracks/<id>/review.md Auto-generated code review report
conductor/tracks/<id>/index.md Track navigation index
conductor/setup_state.json Setup resumption state (last_successful_step)

Cross-session handoff

  • setup_state.json enables resume of interrupted setup from the exact last successful step
  • tracks.md provides cross-session visibility of all work in progress
  • metadata.json per track preserves state for warm-start detection

Context compaction

  • Context is loaded fresh at each command invocation via conductor_cli.py --json calls
  • The CLI consolidates multiple file reads into a single JSON payload, reducing context window consumption
  • batch-match-patterns CLI command replaces per-task pattern calls

Memory persistence scope

  • Project-scoped: all state lives in conductor/ within the repo
  • No global state: each project has its own isolated conductor directory

Audit trail

  • decisions.md per track captures the "why" of implementation choices (ADR logging)
  • review.md per track captures auto-generated code review findings
  • metadata.json tracks status history
07

Orchestration

Claude Conductor (rbarcante) — Orchestration

Multi-agent

Yes — 6 named sub-agents dispatched via Claude's Task tool (and Agent tool in some commands).

  • code-quality-analyzer — dispatched during /conductor:codeReview
  • codebase-pattern-detector — dispatched during brownfield setup
  • git-history-analyst — dispatched during /conductor:revert
  • security-scanner — dispatched during code review
  • test-coverage-analyzer — dispatched during implementation for coverage analysis
  • track-context-researcher — dispatched to consolidate track context before implement

Orchestration pattern

Sequential with conditional sub-agent fan-out. The main track lifecycle is sequential (setup → newTrack → implement → review). Sub-agents are spawned in parallel where independent (e.g., during code review, multiple scanners can run). This is hierarchical: the main command orchestrates, sub-agents do specialized work and report back.

Isolation mechanism

  • Branch per track (via git-isolation.md protocol and suggest-branch CLI)
  • No git worktrees explicitly documented; branches provide logical isolation

Multi-model

No — locked to Claude Code's default model (not configurable per role). model: inherit in frontmatter.

Execution mode

Interactive loop — commands are invoked manually as slash-commands. Not a continuous background daemon. The implement command can chain to newTrack via warm-start but remains user-driven.

Prompt chaining

Yes — newTrack writes spec.md and plan.md, and implement consumes those files as its input. Plan mode output becomes the input for normal-mode execution.

Crash recovery

Yes — setup_state.json enables resume from last successful step. Track metadata.json preserves state so implement can resume.

Context compaction

Handled via CLI consolidation: conductor_cli.py --json tracks read-context TRACK_ID replaces multiple individual file reads, reducing token consumption in long sessions.

08

Ui Cli Surface

Claude Conductor (rbarcante) — UI & CLI Surface

Dedicated CLI

No standalone binary. The conductor_cli.py Python script is invoked internally by the plugin commands, not by users directly.

  • Used by commands via: python ${CLAUDE_PLUGIN_ROOT}/scripts/conductor_cli.py --json <subcommand>
  • Subcommands: newtrack generate-id, newtrack register, implement update-status, implement archive, tracks read-context, implement git-snapshot, implement batch-match-patterns, codereview filtered-diff, tracks update-task, implement modified-files, implement parse-coverage, implement next-adr-number, implement suggest-branch
  • Not exposed as a user-facing binary

Local UI

None — no web dashboard, TUI, or IDE extension.

IDE integration

  • Claude Code plugin only
  • Installed via /plugin marketplace add rbarcante/claude-conductor
  • Slash commands appear as /conductor:<name> in Claude Code

Observability

  • Audit trail via ADRs: decisions.md per track captures architectural decisions
  • Auto code review: review.md per track is auto-generated on track completion
  • Progress tracking: tracks.md and /conductor:status provide status visibility
  • metadata.json: Machine-readable track state

Installation surface area

  • Plugin directory: ~/.claude/plugins/conductor (global) or .claude/plugins/conductor (project-local)
  • Runtime: Python (for CLI helper)
  • No npm, no Node.js required

Related frameworks

same archetype · same primary tool · same memory type

Claude-Flow / Ruflo ★ 55k

Eliminates single-agent context limits and sequential bottlenecks by orchestrating fault-tolerant swarms of specialized AI agents…

Hermes Agent (NousResearch) ★ 168k

Self-improving personal AI agent with closed learning loop, 7 terminal backends, and messaging gateway — not tied to any AI…

OpenCode ★ 165k

Terminal-first AI coding agent with multi-model routing, native desktop app, and a typed .opencode/ configuration system for…

OpenHands ★ 75k

Open-source AI software development platform (open-source Devin alternative) with Docker sandbox isolation, 77.6% SWE-bench…

DeerFlow ★ 70k

Long-horizon superagent that researches, codes, and creates by orchestrating parallel sub-agents with isolated contexts in Docker…

oh-my-openagent (omo) ★ 60k

Multi-provider AI agent orchestration for OpenCode: escape vendor lock-in by routing Sisyphus (Claude/Kimi/GLM) and Hephaestus…